Meta scraps Project OT, underscoring how hard AI deployment really is

What makes a plan like that fragile is the document it is priced against. Lutz Finger, writing in Forbes, argues that adoption stalls because most workflows carry undocumented exceptions, the workarounds everyone quietly performs that never reached a single written procedure ^[6]. A reduction target is a percentage applied to a process description, and the process description is usually incomplete. So the savings are only as real as the mapping work already done, and mapping is the line item that tends to get cut first.

Finger's analogy is the early self-driving car, which learned the rules of the road quickly but could not merge onto a freeway, because people drive by habit and courtesy rather than by rule, and engineers had to sit in the seat and watch until they could describe the behaviour ^[14]. He notes that as the systems matured, the seat needed less engineering and more ordinary judgement ^[20]. The order matters for a budget: the seat gets staffed before it gets emptied.

The adoption figures say slow rather than stalled. The Ifo Institute numbers Finger cites work out to 13.6 percentage points in twelve months, about a third more German firms using AI than a year earlier ^[15]. They also leave roughly 45.5 percent of firms not using it at all ^[16], while on the Harvard Business Review survey he cites, 94 percent of companies stop short of fully trusting agents to run core business processes ^[17]. Those are two different problems. Getting AI into the building is not the same as letting it own an outcome, and money spent on the first does nothing for the second.

The output claims sit on the same fault line. Delivery Hero has said its HeroGen team of coding agents produces output matching roughly 130 developers ^[9]. Finger's question is whether that is more code or more value, and his observation is that AI reliably produces more text and more email while often producing no decisions, because it has no conviction ^[10]. A headcount-equivalent figure counts output, while payroll compensates judgement, and those are not the same currency.

One company's internal reversal could just be internal politics, and a lab chief conceding slow adoption could just be a chief executive managing expectations. Meta has said nothing itself on this record. The check is where hiring money moves. Palantir pioneered the forward deployed engineer, and Finger reports OpenAI, Anthropic, Google, Salesforce and OMMAX now hiring for the role at scale ^[12], with the job defined as embedding inside a client long enough to learn what the org chart and the vendor contract leave out ^[13]. The growth rate comes from one column and no absolute base is given, so it indicates direction rather than size.

The trade-off is unusually clean, and it does not split. The people who hold the undocumented exceptions are generally the people on a reduction list, so a cut booked this quarter removes the input the supervising pods need next quarter. Finger's version of the winning move is building an operating model around AI rather than bolting it onto the existing one ^[19], and that is spending now against savings later. A plan that funds the mappers is slower and duller than one that funds the cut, and it is the one that still exists in six months.